The Agentic State: What’s Next for Digital Government?

All right, thank you. Great to be here and a pleasure to talk you through what we call The Agentic State for the next 20 minutes or so.

So, as already introduced, the work that we do is about governments and what governments can do with technology. I mean, obviously, talk of the town is agentic AI, so we made a lot of thinking around how governments can actually use agentic AI, leading to what we would call The Agentic State. And this is what I'd like to talk you through, in the next 20 minutes.

So, I mean, in very, very basic terms, and I know that this is horribly simplified, but this is what you get if you invite a political scientist to a tech conference that normally speaks to other political scientists. You get sort of simplified things, but still, I think this really matters to understand what agentic AI is and how we work with this, so it's very simple and we say, like, it's pretty much the brain and the hands, right? So the brain, those are the LLMs, other learning models, which are AI, in other words, but then they're applied to workflows.

So you sort of get automatically from A to Z by using the LLMs, up to a situation where you can say that the AI agent, or the software in bigger terms, can perceive or reason or act with minimal human intervention. And this is important to understand, in particular when it comes to workflows, because, look, if we think about what government is, at the end of the day, most of the time, or many times, it's a bundle of workflows, right?

I mean, we all know this. Be it benefit claims, or procurement, or HR licensing, rulemaking, it's really strong bundle of workflows written down, codified workflows in terms of laws, in terms of regulation, and this is actually great terrain for agentic AI, because whenever you have a very clear workflow with clear rules around this, this is a pretty good case to use agentic AI, right? And I mean, secondly, also, why I think we really need this, is that, and that's, I think, an experience that we can all make. Governments oftentimes are slow, reactive.

The way the workflows are treated, executed, is episodic. So also there, I think, there's a big chance and big opportunity, but also a big push that we have to make for governments to use agentic AI. And in particular, this thinking and acting in workflows.

You know, this is something that governments are not really good at. Governments think in terms of political responsibilities, ministries, work charts, and that's sort of incompatible with what we can do with agentic AI, which works on the level of the workflow, right? So a fun exercise, and that's been published a couple of weeks ago, was for us, a report that we did together with the World Economic Forum, to break down governments according to the 70 most important workflows that we name functions. So we said, like, what are the ever-coming-back workflows in governments?

And it turns out that you pretty much can decode a lot of the work done in the public sector according to 70 central workflows. That's been the report. If you wanna read it, there's a QR code. There's also an interactive version on that. And I mean, for everybody who works in government, I think that could be interesting, but also, like, as an interest citizen, please have a look at this, because I think it's really fascinating if we break down government and really understand that at the end of the day, it's not so complex, right? And it's ever and ever-repeating workflows that we have there.

That, as I said before, it's really prime terrain for agentic AI. And on that matter, I mean, just, like, different slide, pretty much the same message. That's recent research by Entropic. You can say, like, what are the things, or what are the different professions that are bound to have a high capability for agentic AI to be used?

Look, I mean, government's not mentioned there exactly, but government, again, is pretty much a mix of office and admin, management, and legal. And those also happen to be precisely those occupations where you can see that there's a big potential for agentic AI. And this was pretty much the thinking that made us really excited to think through what actually this really means to the holistic understanding of government.

And what came out of this, and this is a bit before, and this is something that I would now like to guide you through, is a vision paper that we, so that's me and a couple of further authors, published last October, where we set up the architecture for an agentic state. And, I mean, whoever has something to do with government, if you look at the contributors, it's people like David Knotts, that's the CTO of UK. It's Mikhail Oferov, the Deputy Prime Minister of Ukraine. We've got Sung Woo Kim, the Vice President of the World Bank.

So, you know, that was interesting also for us to understand that there's really a high engagement, and people who, let's be honest, also have other things to do, took the time to contribute to this. While we were writing this, you know, there was also an interesting moment to see, like, this is not only just something happening on paper, but something increasingly also happening in reality.

So, take, for example, the first example there. It's the government of Qatar who builds a building permit, a high-agentic building permit, and cuts down the approval for a building permit down from 60 days to 120 minutes, running live, at scale, live production.

Similarly, I already mentioned him, someone we work closely with is Mikhail Oferov. Back then, when the photo was created, as still the Digital Minister of Ukraine, now the Defense Minister and Deputy Prime Minister of Ukraine, talking through how Ukraine is going to make the move from the digital state to the agentic state.

So, while we were writing the vision paper, we saw that governments are already actively implementing and building parts of the agentic state. And this is the core of the vision paper, and the architecture of the agentic state, as we call it, which has pretty much, I would say, like two different elements. The one is the implementation spectrum. Gonna talk you through that in a second. And then there's also the enablement layers, as we call them. And you also have to read this pretty much in a horizontal manner, but then also in a vertical manner.

So, to guide you through the horizontal layers, what's really important for us to understand, and also to nudge government towards this understanding, is that agentic AI for government is not only about citizen agents. Because this is like the default first answer you often get. It's about citizens that can use agents, and then we sort of have a better user experience for public services. Which is true, of course, and which is something that we cover under public service design and UX. But this is really just one part of what you can do.

What I personally find more interesting, and where I see a bigger potential, is everything related to government workflows. This is the back office of what governments does. High impact, high volume, low complexity, like the permit case from the government of Qatar. Happens round the clock, all the time, in all countries in the world.

So everything that happens in the back office, the government workflows, this is where we see a lot of potential, and we can also see in the vision paper, there's a full chapter on that, outlining what are the dynamics that play out there, and how we can potentially come to a positive scenario for governments using agentic AI there. Where it plays out a bit differently, but also interestingly, is the topic of policy and rulemaking.

Right, I mean, we have a lot of legislative drafting, a lot of scenario mapping that we could do, and with agentic AI you can make this more efficient. But let's also be reminded, many political systems are democracies. Democracies are based on compromise, and on sometimes also bargaining. So it's not always about the best possible solution that you could reach, but also about the human solution that you can reach between people. So this is where I'm a little bit less excited.

Again, big excitement when it comes to regulatory compliance and supervision, because this is, I sense, something that, just take the example, if you had an agent within a company, and you had an agent within the government, and those two agents can communicate and can, for example, say, is the company at all times GDPR compliant? And there's an agent sitting within the company who has access to certain information within the company, and just can then report to the government agent, yes, compliant or non-compliant.

This is, I think, where we can move compliance to something which is real time, and which can be used to really reduce the regulatory burden for both enterprises, firms, but then also for the government. Crisis response, a bit harder, I would say, because you still have the high impact cases, but this is hopefully low volume, right? So you have the black swan events, and there I would say, hard to get the right training data, and hard to have agentic systems that really work that way, but nevertheless, also something to look into.

And then, lastly, public procurements. I don't know if any one of you had ever an interaction point with public procurements, but I am pretty sure that this created headache. It's a long, complicated system with many workflows, both on the buyer's and on the supplier's side, and for a good reason, we have now the term agentic buying, which is more used in the private sector so far, but also can be applied to government. And if you take into consideration that public procurement is about 14% of our GDP, this is also where you can reap a lot of benefits using the agentic AI.

So this is the entire implementation spectrum, and we urge governments to look at this, and say, maybe you start somewhere in the government workflows, maybe you start with public procurement, but really think through that the agentic state is not just one implementation area, but it's those six. And then, as I said, it's really important to work through the enablement layers, because, in particular, the agent governance, this is something that really matters. I think that the speaker coming after me will talk more about this, so I'll keep it short.

Everything about accountability, safety, redress, from a legal perspective, from a process perspective, from a technical perspective, from an organizational perspective, this is what makes up agent governance. And this is what we absolutely need, because if agents start to make decisions on behalf of the state, we have to make sure that we have a system that documents those decisions, and that it's sort of in a bounded autonomy in which those systems can take decisions. Similarly important, data and privacy. I'll skip this a little bit, talk about the tech stack.

We now have more and more, I think, a unified understanding of governments of what's the canonical tech stack. This took time, and I think we are now at a good point for this, but this is a non-agentic AI tech stack, right? So we just, I think, reached the point where governments can sort of say, this is the best practice of how they can build a tech stack, but now it's already time to move forward to build agentic AI capabilities in this tech stack. Gonna come back to this in a minute. Cybersecurity resilience, also very important to build this into the thinking from day one.

Public finance and procurement agents, that's really important for me and really interesting, also because governments are really good at buying person days, but agentic systems don't work by person days, right, but by outcomes. And switching a system which is focused on person days to something that is outcome-based is, within the rules of public procurement, really, really difficult. Nevertheless, important to also think through this proactively. And then lastly, people, culture, and leadership is mainly about that you're gonna have an organization run at two different speeds.

You have the human speed, you have the agentic AI speed, and somehow, you have to bring those two organizations together. Something that happens to all organizations, that's not just something for government, but nevertheless, also important here.

And now, what does it mean for digital identity? Because at the end of the day, that's a conference about digital identity, so let me also walk you through a little bit of the glimpse of where we can see what happens there. I briefly talked about the tech stack.

So, this is how we conceive the tech stack for governments using agentic AI. I'm not gonna go through everything, but just let me point to what's something that we call DPI, short for digital public infrastructure.

This is, by now, the understanding for governments that digital identity, payment, and data sharing mechanisms should be sort of the basic elements that governments use, and this has meant what we call digital public infrastructure. This is designed for a time before agentic AI.

So, there is the urge, indeed, to turn the DPI as we know it towards something that we call the agentic DPI, which is new requirements for digital identity, which is new payment rails, which is new requirements for data, obviously. So, the layer sitting on top of that, for registration interoperability, be it A2A communication or be it MCP service, can actually work with this.

And the reason why I'm telling this is that I think this is really urgent to think about this, and for that, another great study, not by us, Chris Schmitz, a friend, who mapped how government services are increasingly overwhelmed with AI-generated requests. So, look, I mean, those are cases around the world, and just see, for example, that the appeals for social security benefits in Germany, the so-called Bürgergeld, went up almost 100% over a short time. Why is that?

Very easy, everybody now can push everything into JGPT or Claude and say, like, write me an appeal for this and that, and people send it back to governments, right? This is where we stand right now, and this is where we can see that the appeals go up, like, across different countries, across different policy domains. And I think it's really, really easy to understand that we are at the point where people just use GenAI and then put it into government forms, but we are not so far away, right, to say, like, an agent can handle this from me completely autonomously.

So why should I, as a citizen, still make the appeal, right, if the agent's handling the content anyhow, and then I just could leave him and say, so, like, hey, agent, please write me the appeal for Bürgergeld, yeah? For that matter in turn, I need the agent to be able to use my digital identity to be able to then also hand in this appeal. And I think that's gonna happen anywhere, right? And for that matter, it's really important to rethink digital identity not only as something that we can use as citizens, but that potentially also agents that we use in our private life can use on our behalf.

And that's a pretty tricky bit, I think, because right now, full focus in Europe is on the EUDI wallet, rightfully so, but actually the next wave is coming where we have to think proactively through how governments can design digital identities that are not only used by citizens, but also by agents. Let's be honest, the fact that this creates a new big problem for governments, if you get way more appeals, different story.

For that, I think, on the other hand, you have to build the agentic AI capabilities within governments that when the numbers rise by 100% or more, you also have the capabilities to deal with those cases faster. So there's also the push, I would say, from the outside towards government to make the best use of the technology and to be very clear that if the pressure from outside's getting bigger, everybody's using gen AI or then prospectively agentic AI to deal with governments that you also have to have systems internally that can deal also with an agentic suite.

And that being said, thank you very much. That was a bit exotic, I know, but nevertheless, I hope interesting and thank you for the attention. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Great to have you here. Thanks so much for making it. I'm sure everybody really enjoyed that. We have a question here. So who is accountable for decisions an AI agent does in the public sector? What's the state of research into making the reasoning in AI more transparent? First question on the accountability.

The good news is we already sort of have an agentic system within the government, which is a human agentic system. Because everything around accountability, there's a huge pressure for governments, rightfully so, for every civil servant to document decisions, to have audit logs, to have appeal. So for that matter, I think government's really well equipped to build this system for a human agentic system towards a system where also AI agents can act autonomously and accountably.

For example, by audit trails, by logging them decisions, because this is pretty much already a requirement that government has and projecting this towards the agentic AI world. Still gonna be an exercise, but easier in government where I've got clear rules for that than sometimes in the private sector.

Well, thanks very much. Big hand for Manuel Killian. Thank you. Thank you.